Mr. Nilekani, welcome to the Raisina Dialogue. Let me start with a question on the AI race. Now, California is in its own world. It's competing with itself to build the most expensive models, the biggest models, and you've argued that that's a race to the bottom. Is that the right race? Is that the wrong race? And then what's the race to the top?
No, I think there's no doubt that huge amounts of money, a few hundred billion dollars, are being spent on the AI race. But the race I was referring to is what I call a race to the bottom and a race to the top. And what I mean by that is that there are many applications of AI or consequences of AI which are potentially disturbing to society. So the notion of massive job loss because every job gets automated, especially white-collar jobs, knowledge worker jobs, the huge consumption of energy and water for AI as these data centers get built, and that means somebody else has to give up some energy or pay higher prices, the whole addictive nature of the AI tools, the fact that they're addressing things like loneliness for companions, and a whole host of things. And that's happening very fast, and by and large, people are a bit nervous about that, anxious about that. And the race to the top is where AI is being used to do things for society, you know, combat climate change, global hunger, discover new drugs, and so on. So if we want AI to really be successful in the world, we have to make sure there are as many users of it on the positive side as just on the other side. And that's what I meant, that there's a race to the bottom and a race to the top. And currently, I think the race to the bottom is faster than the race to the top.
So I want to stay with the race to the top for a second. India just completed the AI global AI summit, big success. But there's a video of you speaking to the Indian prime minister that has gone viral, and it was a video about building an LLM for Amul, which is the world's largest dairy cooperative. You went from an idea to millions of voices in about 33 days. How did you do it, and is that the race that we should all actually fixate on?
Yeah. So I think what we have learned over the last several years where we tried to build AI is that first of all, technology is only 30% of the problem. 70% is the organizations, the trust building, the coordination amongst various stakeholders, making sure the data is secure, making sure that the answers that you get are not hallucinating. There are a whole host of things like that. And one of the projects we did was called Mahavistar, which is the agricultural bot for farmers in Maharashtra in Marathi and other languages of Maharashtra, and that took about 9 months to build because it required a lot. But that gave us a lot of experience. And then we built something similar for the country of Ethiopia, which was built in 3 months. So we were able to crunch the time from 9 months to 3 months. And when I had this meeting with the prime minister and he brought up the topic of can we do this for cows, because they can't talk back when they are not feeling well, which is a good point. So that's when the PMO connected us to Amul, and we did that in 3 weeks. Now let me explain what's happening there. There are 2.6 million dairy farmers in Amul. There are 22 million cattle in Amul, and there are only 1,400 vets in the entire system. Now if your cow is sick, you don't get access to the vet, or it may take time simply because you have 1,400 vets and 22 million cattle. So what the bot, which has been built called Saraben, which is in Gujarati, which has all the knowledge of all the cows and all the health issues, everything can be used by that woman dairy farmer if she has some challenge with a cow. Now that's a real use. So if we, and already getting thousands of queries, so that's an example of how AI has been put to use for enabling a dairy farmer in Gujarati in the state of Gujarat to get better health care for their cow, which is the principal source of income. To me, that's what we mean by saying that you need to build these kind of things. And what we have learned is now there's a way of building them, but the thing is that technology is only a part of it. It's not just, you know, bring next, it's not LLM, LLM, LLM, it's not like that, it's much more than that.
So just on the LLM, LLM part, I've had many discussions with people here in the morning, in the evening. We had a lively discussion last night, and I was telling them the story about diffusion and AI for people, and they turned it around and saying, oh, that's great, but if you don't have your model, then how are you going to be self-sufficient and resilient in building this out for the future?
Well, I think models are commodities. If you look at what's happening in the model business, in the US alone, there are five huge corporations building large language frontier models. Five. Okay. And they're all spending, and each one is leapfrogging. One day one is in the lead, tomorrow some other. So that's amazing what they're doing. But they're all competing and leapfrogging. And then we have another five Chinese models which are competing almost at the same level. So across the world, we have 10 models in these two countries. In India, we have companies like Sarvam, which has built a 100 billion parameter model, which is a remarkable job with very little, very frugal design, and so on. So they will get built. So when we build a solution like what we have done for Amul, we can use any model we want because the model capacity and power far exceeds what we need in the world of problem solving. They're so far ahead, and this is what we call as the capability gap. The gap between what it can do and what it is doing is so high that there's so much to be done to implement things, to use the power of what we have. If, for example, if not a single dollar was spent on improving these models, we can still use the current models for the next 10 years and really create value for society.
So you've often taught us that technologists don't like geopolitics, but geopolitics loves technologists. And I think a lot of the argument around sovereignty, LLMs, you know, people advising India that you got to build your own model because you can't trust the stuff that the Americans are doing. They argue that you definitely can't trust the stuff that the Chinese are doing. So if I can just push you a little bit, and you often say you're the plumber and you don't have an answer here, but if I can push you, how do we reconcile that? Because that's a reality.
No, I think today in the world of technology, it's always something that can be weaponized, and we have seen that in so many cases. So I think right now, because there are so many choices and the fact we have multiple models, including open-source ones, I think we can solve it. If it comes to a situation where we have to make a decision or something is not working for us, but if we can make it work for us, that's more important. I mean, today we can replace a model in one day, it's not a big deal. So why should we get hung up about that when the real challenge is how do you make it work for millions of farmers or millions of students? That's the hard part.
So I want to come to the hard part. And on models, I think Sam is here, and he's often argued that this is just going to become another utility, which is what you seem to be agreeing with. But then what's the hard part? If this is something that's usable to people, if the model is a commodity, data is of course difficult. What's the bottleneck here for diffusion? Because to make diffusion happen in a society, it deals with millions of users, say farmers. It deals with the government. It deals with weather data. It deals with pricing data from mandis. It deals with warehousing data to know whether there's a warehouse where you can store a few tons of onions or whatever. So this is all over the place. The research, some of the research may come from the ICAR, some may come from another farmer's practices, climate change is happening. How do you bring all these disparate things together in a trusted architecture and deliver consistent answers to the people? That's the hard part.
As diffusion accelerates, do job losses accelerate as well?
That's a separate issue, is how do we think of AI and jobs. And there again, I think this race to the bottom is all, you will not have jobs, and I'm not sure that's a great marketing story to tell, you all, you guys don't have jobs. I mean, what kind of a story is this? So I think the way we are looking at it is actually saying, can we, just like we have a race to the top, can we use AI for job creation or job discovery? And we have this thing called the Blue Dot Initiative, where like when you open your app and you have Ola, Rapido, Uber, whatever, there's a blue dot that shows where the taxi is, you know, and so you can find each other, the consumer can find. Can we do that with jobs? Can a job seeker have a blue dot? Can a job provider have a blue dot? And can we use AI to make them find each other? And we are doing this right now in multiple districts, including in Ghaziabad, where 7,000 people have got jobs because they used the blue dot approach, and the local district commissioner has written about it on his LinkedIn post. So I think again, it goes back to, yes, if you just leave it to the system, then it'll eliminate jobs, or could eliminate, but if you can think of how do we use AI to amplify jobs finding, how to job creation, then there will be. So it goes back to things that this technology is not inevitable, it is something we have to direct in a way that society benefits.
And on directing it, I want to talk a little bit about politics. You are the architect and the creator of India's DPI architecture that I can only imagine back in the day took a lot of politics, and that politics kind of continued in the 15-year journey in themselves. And in some ways, would it be fair to say that politics unleashed the technology on people that then created inclusion and productivity gain? And if you agree with that premise, what does politics need to do on AI? Do politicians understand AI?
No, I think politicians are very wise people, and they understand the power of an idea and how it can benefit people. And I think we have seen that over and over again. I think I talked about my meeting with Prime Minister Modi and talked about Amul. So I think the thing is that people are, that's why this race to the top is important, because for politicians to put their weight behind AI, they should see tangible, visible benefits materially for their people, then they'll put their weight behind AI. The DPI, that's what happened, right? I mean, they saw the power of Aadhaar, they saw the power of UPI, they saw the power of DBT, and you know, put their weight behind it. So it's our job as technologists who see what is possible with some technology to frame it and implement it in a way that society starts seeing benefits in a very inclusive way. I mean, I think politicians will put their weight behind it.
I've got one or two questions, but I'm just going to prep the room. If you have a particular question, start computing them. I'll come to you in about 2 minutes. You've been big on diffusion, and you've just talked to us about the bottlenecks and why it's so important. You've also launched recently a path towards diffusion. I think by 2030 you want 100 pathways. Just open that up for us a little bit.
Well, basically, the idea is that when we did the DPI things, when we took Indian DPI thinking global, we set a target for ourselves. We'll get to 50 countries in 5 years, and we're already at about 39 countries. So we like to be target-oriented. So this time we are saying that if we all believe that AI diffusion is the way to go, can we create 100 different ways of achieving AI diffusion in different sectors? It could be in agriculture, it could be in education, it could be in healthcare. And a diffusion pathway is not technology alone. It's the whole bag of things. How it's implemented, who are the institutions we have to bring on board, what is government policy required, how do we build trusted data, how do we make sure that the sovereignty of data is protected. There's a whole playbook. It's like a playbook for each of them. So we call that as a pathway. And we're assuming if we can do a hundred of them globally, then we'll make a small dent in the argument of how AI is a force for good. And we're hoping that the ones we're doing today in agriculture and education all will multiply. It's really not a technology alone. It's really a whole package of how to do it.
And if I can just take one on the other side of the job loss, which is productivity gain. Where are you on the AI scale? So there's the abundance lot again, the abundance lot in California, there's the frugality lot over here, and then there's the how do you achieve productivity gain. Where's your take on this?
No, I think productivity, definitely AI has a lot of potential of productivity gain, but I think we should distinguish between individual productivity, group productivity, enterprise productivity, and national productivity. So individual productivity is a no-brainer, you know, I can use these tools, Copilot, this and that, and we do all kinds of things, and I can be very pleased with myself. So that's probably very high. But when you get to a group, the productivity drops because there's, you have coordination costs, getting people together. When you take it to an enterprise, it drops even further. And you take it to a society, though, it drops. So I think while there is productivity benefit, we cannot extrapolate from an individual becoming more productive to a society becoming more productive, and that's the fallacy of all this. So I think yes, there will be productivity gains, and definitely things will have a substantial impact, but it'll not be at the pace of how an individual improves.
And last one on government. Governments, do you find generally governments are willing to integrate AI, integrate into workflows and see it as part of their future? Do you see a lot of hesitancy?
Certainly, I think India will embrace AI for the reason that Indians have seen the power of digital in their world. Right? Today we have a woman in a village can get an Aadhaar. With Aadhaar, she can open a bank account, get a mobile connection. She can start a small shop, borrow some money, start selling vegetables or coconuts. She can get a digital history of that with UPI. She can take that history and use the account aggregator system and get a loan and grow her business. So we have created a path of aspiration for a billion people. So they all see the power of technology. So because they see that, they are positively disposed towards technology, and that's what makes India the ideal place to try out all these things to make AI work for people.
AI work for people. Good note to go to people. That's all of you there. Perhaps I could turn to the Raisina fellows to start with. I see a lot of hands up. The Raisina fellows, if anyone, I'm going to privilege you first. So if you have any particular questions, how would you know he's a Raisina fellow? Yeah, they're all in one block, sir. So we've got them in one.
Hello, sir. I am Abhishek Jain. I am the Raisina Young Fellow for this year. So I am a semiconductor professional, and my question is basically on what you said. So as AI models are becoming concentrated in a few global companies or countries like US and China, as you said, how can countries like India avoid becoming merely AI consumers rather than AI creators? Right? So as you know, like in semiconductor, we have lost on the manufacturing in India. Are we risking the same thing with the AI LLM models and technology?
So from consumer to producer, how can we actually, on the back to the models question.
Yeah, so I think the way to think about it is it's about timing. Right now, we should deploy it to make a difference, and that's very hard to do, by the way. It's not easy to pull it off. We have to make that happen. And over time, the models will proliferate, as it is already happening. The models will become cheaper to build, and the knowledge of building models will become ubiquitous. At that time, you build your model. It's not a big deal. Why? It's just a question of sequencing it.
Okay. Thank you. It's a good challenge, I think, for you guys there. Yeah. Next one. Please go ahead.
Good afternoon, sir. My name is Utkash. I'm a Raisina Young Fellow from the '19 cohort. You touched upon productivity at national levels, individual levels, society levels. I was wondering if you had any comments on what happens to achievement, which is far more subjective, and a lot of philosophers say that the sense of achievement with the infusion of AI tends to drop.
